Test generalization across mixture-of-experts model families

Determine whether ResidualMux generalizes across mixture-of-experts model families beyond the single Granite mixture-of-experts model evaluated in the paper.

Background

The paper includes a pilot evaluation on one Granite mixture-of-experts model and reports that norm-normalized injection produces strong receiver recovery while several defenses remain incomplete. These results suggest that residual-norm scaling may be more important than routing alone.

Because only one mixture-of-experts family is evaluated, the reported findings do not establish whether the same vulnerability and defense behavior hold across other mixture-of-experts architectures. The paper explicitly identifies cross-family generalization as untested.